Silicon Before Intelligence: The Hidden Constraint in the Race to AGI

By Shouvik Banerjee

At the recently concluded AI IMPACT Summit in New Delhi, discussions around Artificial General Intelligence (AGI) dominated both policy panels and industry conversations. Much of the debate centred on familiar questions: whether current models can achieve human-level reasoning, how alignment challenges might be addressed, and what governance frameworks are needed for increasingly autonomous systems. The underlying assumption across these discussions was clear: progress toward AGI will primarily be determined by advances in algorithms.

Policy discussions around Artificial General Intelligence (AGI) often focus on software breakthroughs: Model scaling, Reasoning capability, and Model Memory Retention. The implicit assumption is that progress toward advanced AI will be determined primarily by algorithmic innovation. Increasingly, however, evidence suggests that the decisive constraint may lie elsewhere: in semiconductor supply and physical computing infrastructure.

The trajectory of advanced AI is now shaped as much by industrial capacity as by research progress. Memory systems, advanced logic chips, and data centre infrastructure are emerging as binding limits on how quickly AI systems can scale.

AI workloads today differ fundamentally from traditional computing demand. Modern AI models require high-bandwidth memory (HBM) and advanced DRAM architectures capable of sustaining massive parallel processing. Unlike earlier waves of cloud computing, AI training and inference generate sustained pressure on memory bandwidth, interconnect speeds, and energy systems simultaneously. These requirements are pushing semiconductor fabrication capacity toward structural chokepoints rather than temporary shortages.

Industry indicators reinforce this shift. A December 2025 IDC analysis identifies an unprecedented global memory shortage expected to extend into 2027. Supply growth for DRAM and NAND remains below historical trends even as prices rise, driven largely by the rapid expansion of AI data centres.

Because semiconductor fabrication requires years of investment and highly specialised ecosystems, supply cannot rapidly adjust to demand shocks. The result is an increasingly zero-sum allocation dynamic: capacity directed toward AI accelerators and specialised memory reduces availability for consumer electronics and other digital sectors. Rising device costs and slower hardware upgrades therefore reflect a deeper industrial reprioritisation rather than cyclical market behaviour.

Infrastructure pressures extend beyond chips alone. Energy studies project significant increases in electricity consumption driven by AI data centres, while supply-chain research identifies persistent fragility across semiconductor materials, logistics, and fabrication networks. At the same time, export controls and localisation strategies are fragmenting what was once a deeply globalised semiconductor ecosystem. Advanced compute is gradually becoming a strategic asset shaped by geopolitics as much as markets.

This convergence suggests that the current constraint is structural. AI-scale demand, long fabrication lead times, and geopolitical concentration are combining to redefine technological competition around access to compute.

Policy responses are beginning to emerge along three directions.

First, governments are attempting to diversify semiconductor capacity through industrial policy. Initiatives in the United States, Europe, Japan, and India seek to expand domestic fabrication and reduce reliance on geographically concentrated production networks. These efforts reflect a growing recognition that advanced compute infrastructure resembles energy or telecommunications systems — foundational capabilities requiring long-term state coordination rather than purely market-driven investment.

Second, firms are redesigning supply-chain strategies around resilience. Long-term chip procurement agreements, vertical integration, and export-control frameworks increasingly treat AI semiconductors as strategic trade assets. Access to compute is becoming embedded within national competitiveness strategies, reshaping how companies and governments approach technological dependency.

Third, innovation itself is shifting toward efficiency. Architectural changes in AI systems are expected to significantly reduce the cost per token, the core economic metric of AI inference. Techniques that optimise model activation, improve attention mechanisms, and refine training workflows aim to extract greater performance from existing hardware rather than relying solely on larger models. As cost per query declines, progress becomes less dependent on brute-force scaling and more on intelligent utilisation of constrained silicon resources.

This evolution reframes the race toward AGI. Competitive advantage may increasingly depend not only on building more capable models, but on achieving greater intelligence per watt and per unit of memory. Efficiency, once viewed as an engineering optimisation, is becoming a strategic necessity.

For countries such as India, the implication is not immediate technological self-sufficiency but assured participation in emerging compute ecosystems. Investments in semiconductor manufacturing, data centre infrastructure, and energy reliability should therefore be viewed as components of long-term digital sovereignty.

What appears today as a discussion about chip shortages or rising electronics prices is, in reality, an early signal of a broader transformation. Advanced AI development is beginning to encounter material limits that software innovation alone cannot overcome.

The path to AGI will not be determined solely in research laboratories but will also be shaped in fabrication plants, power grids, and supply-chain agreements.

In the emerging AI era, intelligence is increasingly constrained, and enabled, by silicon.

*Disclaimer: Views expressed herein are personal and do not represent those of any organisation or institution

Shouvik Banerjee is a consulting professional with 15+ years of experience in supply chain transformation and AI-enabled operations. His research interests include geopolitics, geoeconomics, and supply chain diplomacy. He is a GCPP-DFA alumnus of the Takshashila Institution.